APIUp to 28% cheaper than official pricesTry the API →
Hermes
Back to News

GitHub's Copilot Now Judges Its Own Code Review Effort Automatically

GitHub announced Copilot code review upgrades at Microsoft Ignite. Adaptive (preview) picks Lite, Balanced, or Max per PR based on complexity and risk, capped by admin limits. Secret detection pairs 500+ regex-based secret types with LLMs and a purpose-built model for unformatted credentials.

GitHub's Copilot Now Judges Its Own Code Review Effort Automatically
Source
GitHub
Published
Author
GitHub
Read
4 min read

Takeaways

  • GitHub announced Copilot code review upgrades at Microsoft Ignite.
  • Adaptive (preview) picks Lite, Balanced, or Max per PR based on complexity and risk, capped by admin limits.
  • Secret detection pairs 500+ regex-based secret types with LLMs and a purpose-built model for unformatted credentials.
  • Cross-repository context (preview, public repos only) finds downstream code broken by your change.
  • Repository insights (GA) tracks addressed comments and merged PRs that skipped Copilot review.
  • Available on Pro, Pro+, Max, Business, and Enterprise; Azure Repos support is now generally available.

GitHub announced five Copilot code review updates at Microsoft Ignite. The reviewer can now adjust its analysis to a pull request’s risk, inspect selected public repositories for related code, and flag possible credentials before merge. GitHub’s changelog covers the three public previews and two generally available features.

FeatureStatusScope
Adaptive review effortPublic previewSelects Lite, Balanced, or Max within an administrator-defined ceiling
Cross-repository contextPublic previewReads selected related public repositories
Secret detectionPublic previewPosts review comments for possible exposed credentials
Repository insightsGenerally availableReports review outcomes for each repository
Azure Repos supportGenerally availableBrings Copilot code review to Azure DevOps repositories

Effort follows risk

Before the Adaptive preview, developers or administrators selected an effort level in settings or when requesting a review. The configuration guide defines three levels:

  • Lite performs a standard review.
  • Balanced examines complex logic, security-sensitive code, and cross-service changes more deeply.
  • Max uses multiple high-reasoning models for pull requests that require the most scrutiny.

Adaptive evaluates each pull request’s complexity and risk, chooses one of those levels, and displays its choice in the review. A small documentation edit may receive a Lite pass, while an authentication change may trigger Max. GitHub has kept the underlying scoring model private.

Administrators retain a hard ceiling because Adaptive cannot select a level above the maximum configured for the enterprise, organization, or repository. Balanced remains the default, so teams must enable Adaptive from the pull request’s Reviewers menu or in personal, repository, organization, or enterprise settings. Azure Repos currently supports Lite and Balanced defaults without Adaptive routing.

Reviews cross repository boundaries

Cross-repository context lets Copilot inspect selected related public repositories during a review. It can trace callers affected by an API change and use documentation, custom instructions, or Copilot skills stored in a central repository. Platform teams can therefore maintain shared review guidance for several services.

GitHub’s demo shows Copilot finding a caller that still passes positional arguments after a function in another repository changed to accept an object. A review limited to the changed repository would lack that dependency context.

The preview only covers public repositories, and teams choose which repositories Copilot may read. Private multi-repository codebases cannot use the feature yet.

Credential checks enter the review

Secret detection combines deterministic pattern matching with model-based analysis. GitHub’s secret scanning partnership program covers more than 500 secret types from over 150 technology partners, using recognizable formats such as vendor-specific prefixes and regular expressions.

A second layer combines large language models with GitHub’s purpose-built secret detection model. It targets credentials without a known format, including internal API keys and database passwords that conventional patterns may miss.

Teams can enable the preview under Settings > Copilot > Code review . The model-based layer can produce false positives, so a limited rollout can establish its precision before organization-wide adoption.

A confirmed credential should be treated as exposed once it has been pushed, even when the pull request never merges. Removing the value in a later commit leaves it in earlier history; the safe response is to revoke or rotate the credential and follow the repository’s history-remediation policy.

Outcomes get a per-repo dashboard

Repository insights gives teams four signals for evaluating Copilot reviews:

  1. Comments addressed
  2. High-severity comments addressed
  3. Pull requests with no later human change request
  4. Merged pull requests that skipped Copilot review

The Copilot usage metrics API exposes the same measures through its repos-1-day reports. Reviews routed by Adaptive are attributed to the effort level it selected.

Comparisons among effort levels should be treated as correlations. Pull request complexity, repository rules, configuration, and team behavior can all affect comment-address rates and later review outcomes.

Azure Repos gets the reviewer

Copilot code review for Azure Repos is generally available for teams using Azure DevOps. Usage is charged as GitHub AI credits to the linked Azure subscription and appears in Azure Cost Management.

GitHub lists Adaptive effort, cross-repository context, and secret detection for Copilot Pro, Pro+, Max, Business, and Enterprise plans. The previews carry no separate feature fee, but Balanced and Max consume more AI credits than Lite. GitHub also notes that higher effort can use slightly more GitHub Actions time.

Adaptive can reduce unnecessary spending by reserving deeper analysis for riskier pull requests. An organization-wide effort ceiling provides a second cost control.

Pilot against a baseline

Teams running Copilot review at scale can evaluate the previews with a controlled rollout:

  1. Record current credit use and Repository insights metrics for a small group of active repositories.
  2. Enable Adaptive with an explicit maximum effort level.
  3. Measure credit use, addressed comments, and later human change requests for several weeks.
  4. Test secret detection separately and classify confirmed findings and false positives.
  5. Enable cross-repository context where public dependencies or shared instructions provide useful review context.

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Sources

  1. 01x.com
  2. 02github.blog
  3. 03github.blog
  4. 04docs.github.com
  5. 05docs.github.com
  6. 06windowsforum.com